In the process of AI product implementation, the quality of the prompt directly determines the effect of the model’s output. This article systematically introduces how to use three structured formats: Markdown, XML, and JSON, to design clear, maintainable, and scalable “production-grade” prompts.
Why Write Structured Prompts?
When prompts become complex, containing multiple instructions, a large amount of context, and specific examples, a flat, unstructured block of text can make it difficult for the model to distinguish between priorities of different information. It’s like giving a long, unfocused verbal instruction to an employee, which can easily lead to misunderstandings and execution deviations.
Structured prompts break down prompts into logically clear modules by using clear separators, labels, or syntax. This brings several core advantages:
- Reduce ambiguity: Clearly indicate which is the instruction, which is the context, and which is the user input, so that the model can accurately understand the intent of each part.
- Improve instruction compliance: When tasks are broken down and presented in a structured manner, the model can better execute step by step, and the output results are more in line with expectations.
- Enhanced maintainabilityStructured prompts are easier for team members to understand, modify, and reuse, reducing long-term maintenance costs.
Why structured formats work: Understand from model training principles
To understand why formats like Markdown, XML, and JSON are particularly effective for LLMs, we need to return to the essence of model training. LLMs work by learning statistical laws in massive amounts of internet text data. A significant portion of this training data is inherently highly structured.
- Data foundation: The internet is full of web pages built in HTML and XML, forum posts and documents written in Markdown, and JSON data widely used in API communication.
- pattern learningDuring the training process, the model has implicitly learned the semantic patterns contained in these structured markers (e.g<h1>., <instruction>”key”: “value”). For example, the model <h1>knows that the text inside the tag is usually the title, the <example>content inside the tag is an example, and the key-value pair structure of JSON represents clear data relationships.
- Activate known modeSo when we use these formats in our prompts, we are not teaching the model a new language, but rather “activating” patterns that it already knows. This allows our instruction intent to be captured more accurately than a normal natural language description. It’s like communicating with someone who is familiar with diagrams, using diagrams is more efficient than plain text descriptions.
Markdown: The first choice for clarity and readability
Markdown is a lightweight markup language popular for its concise and intuitive syntax, making it a starter for writing structured prompts.
B-end product manager’s ability model and learning improvement
The first challenge faced by B-end product managers is how to correctly analyze and diagnose business problems. This is also the most difficult part, product design knowledge is basically not helpful for this part of the work, if you want to do a good job in business analysis and diagnosis, you must have a solid …
View details >
advantage：
- High readability: Use # for title, – or * for list, and text** for bolding, so that the logical hierarchy of the prompt is clear at a glance, and it is very friendly to both humans and AI.
- Easy to write and maintain: The syntax is simple and can be written without complex tools.
- Improve instruction compliance: Research and practice have shown that formatting instructions into bullet points can significantly improve the instruction compliance accuracy of LLMs.
Case study: Restructuring a complex writing task using Markdown
Before refactoring (unstructured prompt):
I need you to write a short story about an astronaut’s discovery of a new planet. The story has to feature a talking alien, a space battle, and an ending reversal that reveals that it’s all just a game. Please keep your story to 500 words.
This prompt mixes all the requirements together, and the model needs to sort out itself.
After refactoring (using Markdown’s structured prompt):
By using Markdown titles and lists, we clearly separate tasks, characters, plots, and constraints, greatly reducing the difficulty of understanding the model, allowing us to generate high-quality content that is more compliant.
Reconstruct prompts with XML: Implement semantic structuring
As a markup language designed for transmitting and storing data, XML (Extensible Markup Language) has core advantages in its self-descriptive and hierarchical structure.
Advantage:
- The semantics are clear: XML’s label structure allows AI to better understand the role and importance of different parts. Compared to plain text, XML provides clear semantic boundaries.
- The layers are clear: Complex prompts often contain multiple layers of information, and the nested structure of XML is a perfect fit for this need.
- Easy to parse: Modern AI models have a good understanding of XML format and can accurately identify and process structured information.
How to write prompt words in XML
XML directives are organized using “tags”, and each instruction part is wrapped in an open label (e.g.<agentinstructions>, ) and a corresponding closed label (</agentinstructions>). The closed label contains a forward slash before the name.
This structure is similar to a box when packing and moving, with the main tab being the largest chest, and subsections like Characters, Goals, and Tasks being smaller boxes nested within them. This layered approach allows for clear organization and modularity, making it easy for both humans and AI to read and process.
Basic structural design principles:
- Use semantic tag names
- Maintain the logic of the hierarchy
- Separate different types of information into different labels
- Use attributes to add metadata
It is recommended to maintain a structured look by indenting nested tags when writing.
JSON: Connect AI with external tools
JSON (JavaScript Object Notation) is a lightweight data exchange format that is mainly used to implement a powerful function in prompt engineering:FunctionCalling。
What is a function call? Function calls are the ability to allow LLMs to interact with external tools or APIs.
When a user’s request requires information outside of the model’s own knowledge base, such as real-time weather, internal database data, or performing a software operation. Instead of answering directly, the model generates a JSON object in a specific format. This JSON object describes exactly which function needs to be called and what parameters need to be passed.
By explicitly requiring LLMs to generate outputs that conform to specific JSONSchemas in prompts, the “last mile” problem of system integration can be solved.
Core strengths
- Reliable machine analysis: The application backend can directly deserialize the JSON string returned by the LLM into an object or data structure without any fragile text parsing logic. This makes the AI’s output as reliable as calling a traditional, deterministic API.
- Seamless system integration: Structured JSON output can be used directly to populate databases, call other APIs, drive front-end UI rendering, or as input to the next step in Agentic Workflows.
- Precise data structure definitions: Defining a JSON schema itself imposes a strong logical constraint on the model, forcing it to think about how to populate the analysis results into preset fields, which in turn improves the quality and completeness of the output.
- Sample data provided: With the JSON format example, the model is able to better understand the desired output format.
Workflow
- Define the tool: In the prompt, use the format of the JSON Schema to describe the available external functions (tools) to the model, including the function name, function description, and parameter list.
- Model decision-making: The model analyzes user input to determine whether a defined function needs to be called to answer the question.
- Generate JSON: If needed, the model outputs a JSON object containing the function name and parameters.
- App execution: Your application code receives this JSON, parses it, and actually executes the corresponding API call or function.
- Return results: Returns the result of the function’s execution to the model.
- Final reply: The model generates a final natural language response to the user based on the results returned by the function.
Example: Implement tool calls through JSON
Step 1: Describe the tool to the model, generallyIn the prompt, we will provide a list of tools.
Explain that this tool is:
- What is it used for?
- What the user input that needs to call this tool looks like
- and which fields are needed to successfully call the tool
Step 2: User Questions User input: “What is Zhang San’s phone number and email address?”
Step 3: Model Generates JSON (Function Call)
Instead of answering directly, the model returns a tool_calls object with the following JSON:
The description calls the personnel query tool, and the field entered is the person’s name Zhang San
Step 4 and beyond: The application code calls the real API and returns the person information
For example, “Zhang San, Tel: 13794785945, Email: zhangsan@123.com”, and then feed it to the model.
The model will finally generate a friendly reply to the user based on the returned information: “Zhang San’s phone is 13794785945, and the email address is zhangsan@123.com”
For product managers and business personnel, understanding function calls is crucial.
It means that LLMs are no longer a closed knowledge base, but “intelligent hubs” that can be authorized to access and operate any external data and systems, which greatly expands the imagination and business value of AI applications.
